Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
Feature Augmentation via Nonparametrics and Selection
(FANS) in High Dimensional Classification∗
Jianqing Fan, Yang Feng, Jiancheng Jiang and Xin Tong
Abstract
We propose a high dimensional classification method that involves nonparametric
feature augmentation. Knowing that marginal density ratios are building blocks of the
Bayes rule for each feature, we use the ratio estimates to transform the original feature
measurements. Subsequently, we invoke penalized logistic regression, taking as input the
newly transformed or augmented features. This procedure trains models with high local
complexity and a simple global structure, thereby avoiding the curse of dimensionality
while creating a flexible nonlinear decision boundary. The resulting method is called
Feature Augmentation via Nonparametrics and Selection (FANS). We motivate FANS
by generalizing the Naive Bayes model, writing the log ratios of joint densities as a
linear combination of those of marginal densities. It is related to generalized additive
models, but has better interpretability and computability. Risk bounds are developed
for FANS. In numerical analysis, FANS is compared with competing methods, so as to
provide a guideline on its best application domain. Real data analysis demonstrates that
FANS performs very competitively on benchmark email spam and gene expression data
sets. Moreover, FANS is implemented by an extremely fast algorithm through parallel
computing.
Keywords: density estimation, classification, high dimensional space, nonlinear decision
boundary, feature augmentation, feature selection, parallel computing.
∗Jianqing Fan is Frederick L. Moore Professor of Finance, Department of Operations Research and
Financial Engineering, Princeton University, Princeton NJ 08544 (Email: [email protected]). Yang
Feng is Assistant Professor, Department of Statistics, Columbia University, New York, NY 10027 (Email:
[email protected]). Jiancheng Jiang is Associate Professor, Department of Mathematics and Statis-
tics, University of North Carolina at Charlotte, Charlotte, NC, 28223 (Email: [email protected]). Xin Tong
is Assistant Professor, Department of Data Sciences and Operations, University of Southern California, Los
Angeles, CA, (Email: [email protected]). The financial support from National Institutes of Health
grants R01-GM072611 and R01GM100474-01 and National Science Foundation grants DMS-1206464 and
DMS-1308566 is greatly acknowledged.
1
1 Introduction
Classification aims to identify to which category a new observation belongs based on feature
measurements. Common applications include disease classification using high-throughput data
such as microarrays, SNPs, spam detection and image recognition. Well known classification
methods include Fisher’s linear discriminant analysis (LDA), logistic regression, Naive Bayes,
k-nearest neighbor, neural networks, and many others. All these methods can perform well
in the classical low dimensional settings, in which the number of features is much smaller
than the sample size. However, in many contemporary applications, the number of features
p is large compared to the sample size n. For instance, the dimensionality p in microarray
data is frequently in thousands or beyond, while the sample size n is typically in the order of
tens. Besides computational issues, the central conflict in high dimensional setup is that the
model complexity is not supported by limited access to data. In other words, the “variance”
of conventional models is high in such new settings, and even simple models such as LDA need
to be regularized. We refer to Hastie et al. (2009) and Buhlmann and van de Geer (2011) for
overviews of statistical challenges associated with high dimensionality.
In this paper, we propose a classification procedure FANS (Feature Augmentation via Non-
parametrics and Selection). Before introducing the algorithm, we first detail its motivation.
Suppose feature measurements and responses are coded by a pair of random variables (X, Y ),
where X ∈ X ⊂ Rp denotes the features and Y ∈ {0, 1} is the binary response. Recall that a
classifier h is a data-dependent mapping from the feature space to the labels. Classifiers are
usually constructed to minimize the risk P (h(X) 6= Y ).
Denote by g and f the class conditional densities respectively for class 0 and class 1, i.e.,
(X|Y = 0) ∼ g and (X|Y = 1) ∼ f . It can be shown that the Bayes rule is 1I(r(x) ≥ 1/2),
where r(x) = E(Y |X = x). Let π = P (Y = 1), then
r(x) =πf(x)
πf(x) + (1− π)g(x).
Assume for simplicity that π = 1/2, then the oracle decision boundary is
{x : f(x)/g(x) = 1} = {x : log f(x)− log g(x) = 0} ,
Denote by g1, · · · , gp the marginals of g, and by f1, · · · , fp those of f . Nonparametric Naive
Bayes model assumes that the conditional distributions of each feature given the class labels
are independent, i.e.,
logf(x)
g(x)=
p∑j=1
logfj(xj)
gj(xj). (1.1)
2
Naive Bayes is a simple approach, but it is useful in high-dimensional settings. Taking a two
class Gaussian model with common covariance matrix, Bickel and Levina (2004) showed that
naively carrying out Fisher’s discriminant rule performs poorly due to diverging spectra. In
addition, the authors argued that independence rule which ignores the covariance structure
performs better than Fisher’s rule in high-dimensional settings. However, correlation among
features is usually an essential characteristic of data, and it can help classification under
suitable models and sample size to feature dimension ratios. Examples in bioinformatics study
can be found in Ackermann and Strimmer (2009). Recently, Fan et al. (2012) showed that
the independence assumption can lead to huge loss in classification power when correlation
prevails, and proposed a Regularized Optimal Affine Discriminant (ROAD). ROAD is a linear
plug-in rule targeting directly on the classification error, and it takes advantages of the un-
regularized pooled sample covariance matrix.
Relaxing the two-class Gaussian assumption in parametric Naive Bayes gives us a general
Naive Bayes formulation such as (1.1). However, this model also fails to capture the corre-
lation, or dependence among features in general. This consideration motivates us to ask the
following question: what if we add weights in front of each log marginal density ratio and
optimize them under certain criterions (all coefficients are 1 in Naive Bayes model, so there
is no need for optimization). More precisely, we would like to learn a decision boundary from
the following set
D =
{x : β0 + β1 log
f1(x1)
g1(x1)+ · · ·+ βp log
fp(xp)
gp(xp)= 0, β0, · · · , βp ∈ R
}. (1.2)
For univariate problems, the marginal density ratio delivers the best classifier based on only
one feature. Therefore, the marginal density ratios can be regarded as the best transformations
of the future measurements, and (1.2) is an effort towards combining those most powerful
univariate transforms to build more powerful classifiers.
This is in a similar spirit as the sure independence screening in Fan and Lv (2008) where
the best marginal predictors are used as probes for their utilities in the joint model. By
combining these marginal density ratios and optimizing over their coefficients βj’s, we wish
to build a good classifier that takes into account feature dependence. Note that although
our target boundary D is not linear in the original features, but it is linear in the parameters
βj’s. Therefore, after renaming the transformed variables, any linear classifiers can be applied.
For example, we can use logistic regression, one of the most popular linear classification rules.
Other choices, such as SVM (linear kernel), are good alternatives, but we fix logistic regression
for the rest of discussion.
3
Recall that logistic regression models the log odds ratio by
logP (Y = 1|X = x)
P (Y = 0|X = x)= β0 +
p∑j=1
βjxj ,
where the βj’s are estimated by the maximum likelihood approach. We should note that with-
out explicitly modeling correlations, logistic regression takes into account features’ joint effects
and levels a good linear combination of features as the decision boundary. Its performance
is similar to LDA, but both models can only capture decision boundaries linear in original
features.
On the other hand, logistic regression might serve as a building block for the more flexible
FANS. Concretely, if we know the marginal densities fj and gj, and run logistic regression
on the transformed features {log(fj(xj)/gj(xj))}, we create a decision boundary nonlinear in
original features. This use of the transformed features is easily interpretable: one naturally
combines the “most powerful” univariate transforms (building blocks of univariate Bayes rules)
{log(fj(xj)/gj(xj))} rather than the original measurements. In special cases such as the
two-class Gaussian model with common covariance matrix, the transformed features are not
different from the original ones. Some caution should be taken: if fj = gj for some j, i.e.,
the marginal densities for feature j are exactly the same, this feature will not have any
contribution in classification. Deletion like this might lose power, because features having no
marginal contribution on their own might boost classification performance when they are used
jointly with other features. In view of this defect, an alternative method is to augment the
original features with the transformed ones.
Since marginal densities fj and gj are unknown, we need to first estimate them, and then
run a penalized logistic regression (PLR) on the estimated transforms. Such regularization
(penalization) is necessary to reduce model complexity in the high dimensional paradigm.
This two-step classification rule of feature augmentation via nonparametrics and selection
will be called “FANS” for short. Precise algorithmic implementation of FANS is described in
the next section. Numerical results show that our new method excels in many scenarios, in
particular when no linear decision boundary can separate the data well.
To understand where FANS stands compared to Naive Bayes (NB), penalized logistic
regression (PLR), and the regularized optimal affine discriminant (ROAD), we showcase a
simple simulation example. In this example, the choice is between a multivariate Gaussian
distribution and the componentwise mixture of two multivariate Gaussian distributions:
Class 0: N((5× 1T10,0Tp−10)T ,Σ),
Class 1: w◦N(0p, Ip)+(1−w)◦N((6×1T10,0Tp−10)T ,Σ), where p = 1000, ◦ is the element-
wise product between matrices, Σii = 1 for all i = 1, · · · , p, Σij = 0.5 for all i, j = 1, · · · , pand i 6= j. w = (w1, · · · , wp)T , where wj ∼iid Bernoulli(0.5).
4
Figure 1: The median test errors for Gaussian vs. mixture of Gaussian when the training
data size varies.
1 1.5 2 2.50
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
log10
(Sample size)
Media
n test cla
ssific
ation e
rror
FANS
ROAD
PLR
NB
The median classification error for 100 repetitions as a function of training sample size n
is rendered in Figure 1. This figure suggests that increasing sample sizes does not help NB
boost performance, because the NB model is severely biased in view of significant correlation
presence. It is interesting to compare PLR with ROAD. ROAD is a more efficient approach
when the sample size is small; however, PLR eventually does better when the sample size
becomes large enough. This is not surprising because the underlying true model is not two
class Gaussian with common covariance matrix. So the less “biased” PLR beats ROAD on
large samples. Nevertheless, even if ROAD uses a misspecified model, it still benefits from a
specific model assumption on small samples. Finally, since the oracle decision boundary in
this example is nonlinear, the newly proposed FANS approach performs significantly better
than others when the sample size is reasonably large. The above analysis seems to suggest
that FANS does well as long as we have enough data to construct accurate marginal density
estimates. Note also that ROAD is better than FANS when the training sample size is
extremely small. The best method in practice largely depends on the available data.
A popular extension of logistic regression and close relative to FANS is the additive lo-
gistic regression, which belongs to generalized additive models (Hastie and Tibshirani, 1990).
5
Additive logistic regression allows (smooth) nonparametric feature transformations to appear
in the decision boundary by modeling
logP (Y = 1|X = x)
P (Y = 0|X = x)=
p∑j=1
hj(xj) , (1.3)
where hj’s are smooth functions. This additive decision boundary is very general, in which
FANS and logistic regression are special cases. It works well for small-p-large-n scenarios,
while its penalized versions adapt to high dimensional settings. We will compare FANS with
penalized additive logistic regression in numerical studies. The major drawback of additive lo-
gistic regression (generalized additive model) is the heavy computational complexity (e.g., the
backfitting algorithm) involved in searching the transformation functions hj(·). Moreover, the
available algorithms, e.g., the algorithm for penGAM (Meier et al., 2009), fail to give an esti-
mate when the sample size is very small. Compared to FANS, the generalized additive model
uses a factor of Kn more parameters, where Kn is the number of knots in the approximation
of every additive components {hj(·)}pj=1. While this reduces possible biases in comparison
with FANS, it increases variances in the estimation and results in more computation cost (see
Table 2). Moreover, FANS admits a nice interpretation of optimal combination of optimal
univariate classifiers.
Besides the aforementioned references, there is a huge literature on high dimensional clas-
sification. Examples include principal component analysis in Bair et al. (2006) and Zou et al.
(2006), partial least squares in Nguyen and Rocke (2002), Huang (2003) and Boulesteix (2004),
and sliced inverse regression in Li (1991) and Antoniadis et al. (2003). Recently, there has
been a surge of interest for extending the linear discriminant analysis to high-dimensional
settings including Guo et al. (2007), Wu et al. (2009), Clemmensen et al. (2011), Shao et al.
(2011), Cai and Liu (2011), Fan et al. (2012), Mai et al. (2012) and Witten and Tibshirani
(2012).
The rest of the paper is organized as follows. Section 2 introduces an efficient algorithm
for FANS. Section 3 is dedicated to simulation studies and real data analysis. Theoretical
results are presented in Section 4. We conclude with a short discussion in Section 5. Longer
proofs and technical results are relegated to the Appendix.
2 Algorithm
In this section, we detail an efficient algorithm (S1 − S5) for FANS. A variant of FANS
(FANS2), which uses the transformed features to augment the original ones, is also described.
6
2.1 FANS and its Running Time Bound
S1. Given n pairs of observations D = {(X i, Yi), i = 1, · · · , n}. Randomly split the data
into two parts for L times: Dl = (Dl1, Dl2), l = 1, · · · , L.
S2. On each Dl1, l ∈ {1, · · · , L}, apply kernel density estimation and denote the estimates
by f = (f1, · · · , fp)T and g = (g1, · · · , gp)T .
S3. Calculate the transformed observations Zi = Z f ,g(X i), where Zij = log fj(Xij) −log gj(Xij), for each i ∈ Dl2 and j ∈ {1, . . . , p}.
S4. Fit an L1-penalized logistic regression to the transformed data {(Zi, Yi), i ∈ Dl2}, using
cross validation to get the best penalty parameter. For a new observation x, we estimate
transformed features by log fj(xj)−log gj(xj), j = 1, . . . , p, and plug them into the fitted
logistic function to get the predicted probability pl = P (Y = 1|X = x).
S5. Repeat (S2)-(S4) for l = 1, · · · , L, use the average predicted probability prob = L−1∑L
l=1 pl
as the final prediction, and assign the observation x to class 1 if prob ≥ 1/2, and 0 oth-
erwise.
A few comments on the technical implementation are made as follows.
Remark 1
i). In S2, if an estimated marginal density is less than some threshold ε (say 10−2), we set
it to be ε. This Winsorization increases the stability of the transformations, because the
estimated transformations log fj and log gj are unstable in regions where true densities
are small.
ii). In S4, we take penalized logistic regression, but any linear classifier can be used. For
example, support vector machine (SVM) with linear kernel is also a good choice.
iii). In S4, the L1 penalty (Tibshirani, 1996) was adopted since our primary interest is the
classification error. We can also apply other penalty functions, such as SCAD (Fan and
Li, 2001), adaptive LASSO (Zou, 2006) and MCP (Zhang, 2010).
iv). In S5, the average predicted probability is taken as the final prediction. An alternative
approach is to make a decision on each random split, and listen to majority votes.
In S1, we split the data multiple times. The rationale behind multiple splitting lies in
the two-step prototype nature of FANS, which uses the first part of the data for marginal
7
nonparametric density estimates (in S2) and (transformation of) the second part for penal-
ized logistic regression (in S4). Multiple splitting and prediction averaging not only make
our procedure more robust against arbitrary assignments of data usage, but also make more
efficient use of limited data. We take L = 20 in our numerical studies. This choice reflects
our cluster’s node number. Interested readers can as well leverage their better computing
resources for a larger L, but our experience is that more splits do not help universally in our
numerical examples. Also, we recommend a balanced assignment by switching the role of data
used for feature transformation and for feature selection, i.e., D2l = (D(2l−1),2, D(2l−1),1) when
D2l−1 = (D(2l−1),1, D(2l−1),2).
It is straightforward to derive a running time bound for our algorithm. Suppose splitting
has been done. In S2, we need to perform kernel density estimation for each variable, which
costs O(n2p)1. The transformations in S3 cost O(np). In S4, we call the R package glmnet
to implement penalized logistic regression, which employs the coordinate decent algorithm
for each penalty level. This step has a computational cost at most O(npT ), where T is the
number of penalty levels, i.e., the number of times the coordinate descent algorithm is run
(see Friedman et al. (2007) for a detailed analysis). The default setting is T = 100, though
we can set it to other constants. Therefore, a running time bound for the whole algorithm is
O(L(n2p+ np+ npT )) = O(Lnp(n+ T )).
The above bound does not look particularly interesting. However, smart implementation
of the FANS procedure can fully unleash the potential of our algorithm. Indeed, not only
the L repetitions, but also the marginal density estimates in S2 can be done via parallel
computing. Suppose L is the number of available nodes, and the cpu core number in each
node is N ≥ n/T . This assumption is reasonable because T = 100 by default, N = 8 for
our implementation, and sample sizes n for many applications are less than a multiple of
TN . Under this assumption, the L predicted probabilities calculations can be carried out
simultaneously and the results are combined later in S5. Moreover in S2, the running time
bound becomes O(n2p/N). Henceforth, a bound for the whole algorithm will be O(npT ),
which is the same as for penalized logistic regression. The exciting message here is that,
by leveraging modern computer architecture, we are able to implement our nonparametric
classification rule FANS within running time at the order of a parametric method. The
computation times for various simulation setups are reported in Table 2, where the first
column reports results when only L repetitions are paralleled, and the second column reports
the improvement when marginal density estimates in S2 are paralleled within each node.
1Approximate kernel density estimates can be computed faster, see e.g., Raykar et al. (2010).
8
2.2 Augmenting Linear Features
As we argued in the introduction, features with no marginal discrimination power do not make
contribution in FANS. One remedy is to run (in S4) the penalized logistic regression using
both the transformed features and the original features, which amounts to modeling the log
odds ratio by
β0 + β1 logf1(x1)
g1(x1)+ . . .+ βp log
fp(xp)
gp(xp)+ βp+1x1 + . . .+ β2pxp .
This variant of FANS is named FANS2, and it allows features with no marginal power to enter
the model in a linear fashion. FANS2 augments the original features with the log-likelihood
ratios and helps when a linear decision boundary separates data reasonably well.
3 Numerical Studies
3.1 Simulation
In simulation studies, FANS and FANS2 are compared with competing methods: penalized
logistic regression (PLR, Friedman et al. (2010)), penalized additive logistic regression models
(penGAM, Meier et al. (2009)), support vector machine (SVM), regularized optimal affine
discriminant (ROAD, Fan et al. (2012)), linear discriminant analysis (LDA), Naive Bayes
(NB) and feature annealed independence rule (FAIR, Fan and Fan (2008)).
In all simulation settings, we set p = 1000 and training and testing data sample sizes
of each class to be 300. Five-fold cross-validation is conducted when needed, and we repeat
50 times for each setting (The relative small number of replications is due to the intensive
computation time of penGAM, c.f. Table 2). Table 1 summarizes median test errors for each
method along with the corresponding standard errors. This table omits Fisher’s classifier
(using pseudo inverse for sample covariance matrix), because it gives a test error around 50%,
equivalent to random guessing.
Example 1 We consider the two class Gaussian settings where Σii = 1 for all i = 1, · · · , pand Σij = ρ|i−j|, µ1 = 01000 and µ2 = (1T10,0
T990)T , in which 1d is a length d vector with all
entries 1, and 0d is a length d vector with all entries 0. Two different correlations ρ = 0 and
ρ = 0.5 are investigated.
This is the classical LDA setting. In view of the linear optimal decision boundary, the
nonparametric transformations in FANS is not necessary. Table 1 indicates some efficiency
(not much) loss due to the more complex model FANS. However, by including the original
features, FANS2 is comparable to the methods (e.g., PLR and ROAD) which learn linear
9
boundaries of original features. In other words, the price to pay for using the unnecessarily
more complex method FANS (FANS2) is small in terms of the classification error.
An interesting observation is that penGAM, which is based on a more general model class
than FANS and FANS2, performs worse than our new methods. This is also expected as
the complex parameter space considered by penGAM is unnecessary in view of linear optimal
decision boundary. Surprisingly, SVM performs poorly (even worse than NB), especially when
all features are independent.
Example 2 The same settings as Example 1 except the common covariance matrix is an equal
correlation matrix, with a common correlation ρ = 0.9.
Same as in Example 1, FANS and FANS2 have performance comparable to PLR and
ROAD. Although FAIR works very well in Example 1, where the features are independent (or
nearly independent), it fails badly when there is significant global pairwise correlation. Similar
observations also hold for NB. This example shows, ignoring correlation among features could
lead to significant loss of information and deterioration in the classification error.
Example 3 One class follows a multivariate Gaussian distribution, and the other a mixture
of two multivariate Gaussian distributions. Precisely,
Class 0: N((3× 1T10,0Tp−10)T ,Σp),
Class 1: 0.5×N(0p, Ip) + 0.5×N((6× 1T10,0Tp−10)T ,Σp),
where Σii = 1, Σij = ρ for i 6= j. Correlations ρ = 0 and ρ = 0.5 are considered.
In this example, Class 0 and Class 1 have the same mean, so the differences are in higher
order moments. Table 1 shows that all methods based on linear boundary perform like random
guessing, because the optimal decision boundary is highly nonlinear. penGAM is comparable
to FANS and FANS2, but SVM cannot capture the oracle decision boundary well even if a
nonlinear kernel is applied.
Example 4 Two classes follow uniform distributions,
Class 0: Unif (A),
Class 1: Unif (B\A),
where A = {x ∈ Rp : ‖x‖2 ≤ 1} and B = [−1, 1]p.
Clearly, the oracle decision boundary is {x ∈ Rp : ‖x‖2 = 1}. Again, FANS and FANS2
capture this simple boundary well while the linear-boundary based methods fail to do so.
Computation times (in seconds) for various classification algorithms are reported in Table
2. FANS is extremely fast thanks to parallel computing. While penGAM performs similarly
to FANS in the simulation examples, its computation cost is much higher. The similarity
10
in performance is due to the abundance in training examples. We will demonstrate with an
email spam classification example that penGAM fails to deliver satisfactory results on small
samples.
Table 1: Median test error (in percentage) for the simulation examples. Standard errors are
in the parentheses.
Ex(ρ) FANS FANS2 ROAD PLR penGAM NB FAIR SVM
1(0) 6.8(1.1) 6.2(1.2) 6.0(1.3) 6.5(1.2) 6.6(1.1) 11.2(1.4) 5.7(1.0) 13.2(1.5)
1(0.5) 16.5(1.7) 16.2(1.8) 16.5(5.3) 15.9(1.7) 16.9(1.6) 20.6(1.7) 17.2(1.6) 22.5(1.8)
2(0.5) 4.2(0.9) 2.0(0.6) 2.0(0.6) 2.5(0.6) 3.7(0.9) 43.5(11.1) 25.3(1.6) 5.3(1.1)
2(0.9) 3.1(1.1) 0.0(0.0) 0.0(0.0) 0.0(0.0) 0.2(1.4) 46.8(8.8) 30.2(1.9) 0.0(0.1)
3(0) 0.0(0.0) 0.0(0.0) 49.6(2.4) 50.0(1.3) 0.0(0.1) 50.4(2.2) 50.2(2.1) 31.8(2.4)
3(0.5) 3.4(0.7) 3.4(0.7) 49.3(2.4) 50.0(1.3) 3.7(0.8) 50.0(2.1) 50.2(2.0) 19.8(2.4)
4 0.0(0.0) 0.0(0.0) 28.2(1.8) 50.0(10.7) 0.0(0.0) 41.0(1.1) 34.6(1.4) 0.0(0.0)
Table 2: Computation time (in seconds) comparison for FANS, SVM, ROAD and penGAM.
The parallel computing technique is applied. Standard errors are in the parentheses.
Ex(ρ) FANS FANS(para) SVM ROAD penGAM
1(0) 12.0(2.6) 3.8(0.2) 59.4(12.8) 99.1(98.2) 243.7(151.8)
1(0.5) 12.7(2.1) 3.5(0.2) 81.3(19.2) 100.7(89.3) 325.8(194.3)
2(0.5) 16.0(3.1) 4.0(0.2) 77.6(18.1) 106.8(90.7) 978.0(685.7)
2(0.9) 22.0(4.6) 4.5(0.3) 75.7(17.8) 98.3(83.9) 3451.1(3040.2)
3(0) 12.1(2.1) 3.4(0.2) 152.1(27.4) 96.3(68.8) 254.6(130.0)
3(0.5) 11.9(2.0) 3.4(0.2) 342.1(58.0) 95.9(74.8) 298.7(167.4)
4 22.4(3.9) 6.6(0.4) 264.3(45.0) 75.1(54.0) 4811.9(3991.7)
3.2 Real Data Analysis
We study two real examples, and compare FANS(FANS2) with competing methods.
11
3.2.1 Email Spam Classification
First, we investigate a benchmark email spam data set. This data set has been studied by
Hastie et al. (2009) among others to demonstrate the power of additive logistic regression
models. There are a total of n = 4, 601 observations with p = 57 numeric attributes. The
attributes are, for instance, the percentage of specific words or characters in an email, the
average and maximum run lengths of upper case letters, and the total number of such letters.
To show suitable application domains of FANS and FANS2, we vary the training proportion,
from 5%, 10%, 20%, · · · , to 80% of the data while assigning the rest as test set. Splits are
repeated for 100 times and we report the median classification errors.
Figure 2 and Table 3 summarize the results. First, we notice that FANS and FANS2
are very competitive when training sample sizes are small. As the training sample size in-
creases, SVM becomes comparable to FANS2 and slightly better than FANS. In general, these
three methods dominate throughout different training proportions. The more complex model
penGAM failed to yield classifiers when training data proportion is less than 30% due to the
difficulty of matrix inversion with the splines basis functions. For larger training samples,
penGAM performs better than linear decision rules; however, it is not as competitive as either
FANS or FANS2. Also interestingly, Naive Bayes (NB) is the favored method given smallest
training sample (5%), and but its test error remains almost unchanged when the sample size
increases. In other words, NB’s independence assumption allows good training given very few
data points, but it cannot benefit from larger samples due to severe model bias.
Table 3: Median classification error (in percentage) on e-mail spam data when the size of the
training data varies. Standard errors are in the parentheses.
% FANS FANS2 ROAD PLR penGAM LDA NB FAIR SVM
5 11.1(2.6) 10.5(1.1) 13.6(0.9) 13.5(1.7) - 13.6(1.1) 10.5(5.0) 15.6(1.7) 11.2(0.8)
10 8.7(2.4) 8.5(0.9) 11.3(0.8) 10.5(1.1) - 11.3(0.9) 10.7(4.2) 13.5(0.9) 9.4(0.7)
20 8.0(2.1) 7.7(0.7) 10.6(0.6) 9.0(0.8) - 10.3(0.6) 10.7(5.3) 12.4(0.7) 8.1(0.7)
30 7.8(1.7) 7.4(0.5) 10.3(0.4) 8.9(0.6) 9.2(0.6) 10.1(0.5) 10.7(4.0) 11.7(0.4) 7.4(0.6)
40 7.2(2.2) 6.9(0.5) 10.1(0.5) 9.0(0.6) 8.6(0.5) 10.0(0.4) 10.5(5.1) 11.5(0.6) 7.0(0.5)
50 7.4(2.2) 7.0(0.5) 9.9(0.5) 8.5(0.6) 8.3(0.5) 9.9(0.4) 10.7(4.1) 11.8(0.6) 6.9(0.5)
60 7.4(2.2) 6.8(0.5) 9.8(0.6) 9.3(0.6) 7.8(0.6) 9.5(0.5) 10.6(4.8) 11.8(0.7) 6.5(0.6)
70 7.2(1.6) 6.4(0.6) 9.5(0.7) 9.2(0.7) 7.4(0.7) 9.4(0.6) 10.5(4.6) 11.4(0.7) 6.4(0.7)
80 6.9(1.6) 6.3(0.7) 9.4(0.6) 9.3(0.9) 7.4(0.8) 9.2(0.6) 10.4(4.7) 11.4(0.8) 6.3(0.9)
12
Figure 2: The median test classification error for the spam data set using various proportions
of the data as training sets for different classification methods.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.80.06
0.07
0.08
0.09
0.1
0.11
0.12
0.13
0.14
0.15
0.16
Proportion of Training Data
Media
n T
est C
lassific
ation E
rror
PLR
FANS
FANS2
SVM
ROAD
LDA
NB
FAIR
penGAM
Table 4: Classification error and number of selected genes on lung cancer data.
FANS FANS2 ROAD PLR penGAM FAIR NB
Training Error 0 0 1 0 0 0 6
Testing Error 0 0 1 6 2 7 36
No. of selected genes 52 52 52 15 16 54 12533
3.2.2 Lung Cancer Classification
We now evaluate the newly proposed classifiers on a popular gene expression data set “Lung
Cancer” (Gordon et al., 2002), which comes with predetermined, separate training and test
sets. It contains p = 12, 533 genes for n0 = 16 adenocarcinoma (ADCA) and n1 = 16
mesothelioma training vectors, along with 134 ADCA and 15 mesothelioma test vectors.
Following Dudoit et al. (2002), Fan and Fan (2008), and Fan et al. (2012), we standardized
each sample to zero mean and unit variance. The classification results for FANS, FANS2,
ROAD, penGAM, FAIR and NB are summerized in Table 4. FANS and FANS2 achieve 0 test
classification error, while the other methods fail to do so.
13
4 Theoretical Results
In this section, an oracle inequality regarding the excess risk is derived for FANS. Denote
by f = (f1, · · · , fp)T and g = (g1, · · · , gp)T vectors of marginal densities of each class with
f 0 = (f0,1, · · · , f0,p)T and g0 = (g0,1, · · · , g0,p)
T being the true densities. Let {(X i, Yi)}ni=1 be
i.i.d. copies of (X, Y ), and the regression function be modeled by
P (Y1 = 1|X1) =1
1 + exp(−m(Z1)),
where Z1 = (Z11, · · · , Z1p)T , each Z1j = Z1j(X1) = log fj(X1j) − log gj(X1j), and m(·) is a
generic function in some function class M including the linear functions. Now, let Q = {q =
(m,f , g)} be the parameter space of interest with constraints on m, f and g be specified
later. The loss function we consider is
ρ(q) = ρ(m,f , g) = ρq(X1, Y1) = −Y1m(Z1) + log(1 + exp[m(Z1)]) .
Let m0 = arg minm∈M Pρ(m,f 0, g0). Then the target parameter is q∗ = (m0,f 0, g0). We use
a working model with mβ(Z1) = βTZ1 to approximate m0. Under this working model, for a
given parameter q = (mβ,f , g), let
πq(X1) = P (Y1 = 1|X1) =1
1 + exp(−βTZ1).
With this linear approximation, the loss function is the logistic loss
ρ(q) = ρq(X1, Y1) = −Y1βTZ1 + log(1 + exp[βTZ1]) .
Denote the empirical loss by Pnρ(q) =∑n
i=1 ρq(X i, Yi)/n, and the expected loss by Pρ(q) =
Eρq(X, Y ). In the following, we take M as linear combinations of the transformed features
so that m0 = mβ0, where
β0 = arg minβ∈Rp
Pρ(mβ,f 0, g0) .
In other words, q0 = (mβ0,f 0, g0) = q∗. Hence, the excess risk for a parameter q is
E(q) = P [ρ(q)− ρ(q∗)] = P [ρ(q)− ρ(q0)] . (4.4)
As described in Section 2, densities f 0 and g0 are unavailable and must be estimated. Dif-
ferent from the numerical implementation, we now proceed with a modified sampling scheme
and procedure that are more friendly for theoretical derivation. Suppose we have labeled
samples {X+1 , · · · ,X+
n1} (used to learn f 0) and {X−1 , · · · ,X−n1
} (used to learn g0; theory
carries over for different sample sizes), in addition to an i.i.d. sample {(X1, Y1), · · · , (Xn, Yn)}
14
(used to select features). Moreover, suppose {(X1, Y1), · · · , (Xn, Yn)} is independent of
{X+1 , · · · ,X+
n1} and {X−1 , · · · ,X−n1
}. A simple way to comprehend the above theoretical
set up is that the sample size of 2n1 +n has been split into three groups. The notations P and
E are regarding the random couple (X, Y ). We use the notation P n to denote the probability
measure induced by the sample {(X1, Y1), · · · , (Xn, Yn)}, and notation P+ and P− for the
probability measure induced by the labeled samples {X+1 , · · · ,X+
n1} and {X−1 , · · · ,X−n1
}.The density estimates f = (f1, · · · , fp)T and g = (g1, · · · , gp)T are based on samples
{X+1 , · · · ,X+
n1} and {X−1 , · · · ,X−n1
}:
fj(x) =1
n1h
n1∑i=1
K
(X+ij − xh
)and gj(x) =
1
n1h
n1∑i=1
K
(X−ij − x
h
)for j = 1, · · · , p ,
in which K(·) is a kernel function and h is the bandwidth. Then with these estimated marginal
densities, we have an “oracle estimate” q1 = (β1, f , g), where
β1 = arg minβ∈Rp
Pρ(mβ, f , g) .
It is the oracle given marginal density estimates f and g, and is estimated in FANS by
β1 = arg minβ∈Rp
Pnρ(mβ, f , g) + λ‖β‖1 .
Let q1 = (mβ1, f , g). Our goal is to control the excess risk E(q1), where E is defined by (4.4).
In the following, we introduce technical conditions for this task.
Let Z0 be the n × p design matrix consisting of transformed covariates based on the
true densities f 0 and g0. That is Z0ij = log f0,j(Xij) − log g0,j(Xij), for i = 1, · · · , n and
j = 1, · · · , p. In addition, let Z0 = (Z01,Z
02, . . . ,Z
0n)T . Also, denote by |S| the cardinality of
the set S. Denote by ‖D‖max = maxij |Dij| for any matrix D with elements Dij.
Assumption 1 (Compatibility Condition) The matrix Z0 satisfies compatibility condi-
tion with a compatibility constant φ(·), if for every subset S ⊂ {1, · · · , p}, there exists a
constant φ(S), such that for all β ∈ Rp that satisfy ‖βSc‖1 ≤ 3‖βS‖1, it holds that
‖βS‖21 ≤
1
nφ2(S)‖Z0β‖2|S| .
A direct application of Corollary 6.8 in Buhlmann and van de Geer (2011) leads to a compati-
bility condition on the estimated transform matrix Z, in which Zij = log fj(Xij)− log gj(Xij).
Lemma 1 Denote by E = Z − Z0 the estimation error matrix of Z0. If the compatibility
condition is satisfied for Z0 with a compatibility constant φ(·), and the following inequalities
hold
32‖E‖max|S|φ(S)2
≤ 1 , for every S ⊂ {1, · · · , p}, (4.5)
the compatibility condition holds for Z with a new compatibility constant φ1(·) ≥ φ(·)/√
2.
15
The Compatibility Condition can be interpreted as a condition that bounds the restricted
eigenvalues. The irrepresentable condition (Zhao and Yu, 2006) and the Sparse Riesz Con-
dition (SRC) (Zhang and Huang, 2008) are in similar spirits. Essentially, these conditions
avoid high correlation among subsets where signals are concentrated; such high correlation
may cause difficulty in parameter estimation and risk prediction.
To help theoretical derivation, we introduce two intermediate L0-penalized estimates.
Given the true densities f 0 and g0, consider a penalized theoretical solution q∗0 = (β∗0,f 0, g0),
where
β∗0 = arg minβ∈Rp
3Pρ(mβ,f 0, g0) + 2H
(4λ√sβ
φ(Sβ)
), (4.6)
in which H(·) is a strictly convex function on [0,∞) with H(0) = 0, sβ = |Sβ| is the cardinality
of Sβ = {j : βj 6= 0}, and φ(·) is the compatibility constant for Z0. Throughout the paper,
we consider a specific quadratic function2 H(v) = v2/(4c) whose convex conjugate is G(u) =
supv{uv −H(v)} = cu2. Then, equation (4.6) defines an L0-penalized oracle:
β∗0 = arg minβ∈Rp
3Pρ(mβ,f 0, g0) +8λ2sβcφ2(Sβ)
. (4.7)
Similarly, with density estimate vectors f and g, we define an L0-penalized oracle estimate
q∗1 = (mβ∗1, f , g), where
β∗1 = arg minβ∈Rp
3Pρ(mβ, f , g) +8λ2sβcφ2
1(Sβ). (4.8)
To study the excess risk E(q1), we consider its relationship with E(q∗1) and E(q∗0).
Assumption 2 (Uniform Margin Condition) There exists η > 0 such that for all (mβ,f , g)
satisfying ‖β − β0‖∞ + max1≤j≤p ‖fj − f0,j‖∞ + max1≤j≤p ‖gj − g0,j‖∞ ≤ 2η, we have
E(mβ,f , g) ≥ c‖β − β0‖22 , (4.9)
where c is the positive constant in (4.7).
The uniform margin condition is related to the one defined in Tsybakov (2004) and van de
Geer (2008). It is a type of “identifiability” condition. Basically, near the target parameter
q0 = (mβ0,f 0, g0), the functional value needs to be sufficiently different from the value on
q0 to enable enough separability of parameters. Note that we impose the uniform margin
condition in both the neighborhood of the parametric component β0 and the nonparametric
2The following theoretical results can be derived for a generic strictly convex function H(·) along the same
lines.
16
components f 0 and g0, because we need to estimate the densities, in addition to the parametric
part. A related concept in binary classification is called “Margin Assumption”, which was first
introduced in Polonik (1995) for densities.
To study the relationship between E(q1) and E(q∗1), we define
vn(β) = (Pn − P )ρ(mβ, f , g) and WM = sup‖β−β∗
1‖≤M|vn(β)− vn(β∗1)| .
Denote by
2ε∗ = 3E(mβ∗1, f , g) +
8λ2sβ∗1
cφ21(Sβ∗
1).
Set M∗ = ε∗/λ0 (λ0 to specified in Theorem 1) and
J1 = {WM∗ ≤ λ0M∗} = {WM∗ ≤ ε∗} .
The idea here is to choose λ0 such that the event J1 has high probability.
A few more notations are introduced to facilitate the discussion. Let τ > 0. Denote by
bτc the largest integer strictly less than τ . For any x, x′ ∈ R and any bτc times continuously
differentiable real valued function u on R, we denote by ux its Taylor polynomial of degree
bτc at point x:
ux(x′) =
∑|s|≤bτc
(x′ − x)s
s!Dsu(x) .
For L > 0, the (τ, L, [−1, 1])-Holder class of functions, denoted by Σ(τ, L, [−1, 1]), is the
set of functions u : R → R that are bτc times continuously differentiable and satisfy, for any
x, x′ ∈ [−1, 1], the inequality:
|u(x′)− ux(x′)| ≤ L|x− x′|τ .
The (τ, L, [−1, 1])-Holder class of density is defined as
PΣ(τ, L, [−1, 1]) =
{p : p ≥ 0,
∫p = 1, p ∈ Σ(τ, L, [−1, 1])
}.
Assumption 3 Assume that β1 is in the interior of some compact set Cp. There exists an
ε0 ∈ (0, 1) such that for all β ∈ Cp, f , g ∈ PΣ(2, L, [−1, 1]), ε0 < π(mβ,f ,g)(·) < 1− ε0.
Assumption 4 ‖Z0‖max ≤ K for some absolute constant K > 0, and ‖β0‖∞ ≤ C1 for some
absolute constant C1 > 0.
Assumption 5 The penalty level λ is in the range of (8λ0, Lλ0) for some L > 8. Moreover,
the following holds
8KL2(eη/ε0 + 1)2
η
λ0sβ∗1
φ21(Sβ∗
1)≤ 1,
where η is as in the uniform margin condition.
17
Assumption 3 is a regularity condition on the probability of the event that the observation
belongs to class 1. Since the FANS estimator is based on the estimated densities, we impose
the constraints in a neighborhood of the oracle estimate β1 (when using f and g). Assumption
4 bounds the maximum absolute entry of the design matrix as well as the maximum absolute
true regression coefficient. Assumption 5 posits a proper range of the penalty parameter λ to
guarantee that the penalized estimator mimics the un-penalized oracle.
Assumption 6 Suppose the feature measurement X has a compact support [−1, 1]p, and
f0,j, g0,j ∈ PΣ(2, L, [−1, 1]) for all j = 1, · · · , p, where PΣ denotes a Holder class of densities.
Assumption 7 Suppose there exists εl > 0 such that for all j = 1, · · · , p, εl ≤ f0,j, g0,j ≤ ε−1l .
Also we truncate estimates fj and gj at εl and ε−1l .
Assumption 8
n720− 3
4α
1 (log(3p))34 (log n1)
110 = o(1) ,
and,
n110−α
1 (log(3p))12 (log n1)
25 = o(1) ,
for some constant α > 7/15.
Assumption 6 imposes constraints on the support of X and smoothness condition on the
true densities f 0 and g0, which help control the estimation error incurred by the nonparametric
density estimates. Assumption 7 assumes that the marginal densities and the kernel are strictly
positive on [−1, 1]p. Assumption 8 puts a restriction on the growth of the dimensionality p in
terms of sample size n1.
We now provide a lemma to bound the uniform deviation between fj and f0,j for j =
1, · · · , p.
Lemma 2 Under Assumptions 6-8, taking the bandwidth h =(
logn1
n1
)1/5
, for any δ1 > 0,
there exists N∗1 such that if n1 ≥ N∗1 ,
P+−(
max1≤j≤p
‖fj − f0,j‖∞ ≥ m
)≤ δ1 , and P+−
(max1≤j≤p
‖gj − g0,j‖∞ ≥ m
)≤ δ1 ,
for m = C2
√2 log(3p/δ1)
n1−α1
, and C2 is an absolute constant.
Denote by
J2 =
{max1≤j≤p
‖fj − f0,j‖∞ ≤ η/2, max1≤j≤p
‖gj − g0,j‖∞ ≤ η/2
},
where η is the constant in the uniform margin condition. It is straightforward from Lemma 2
that
P+−(J2) ≥ 1− 6p
exp(η2n1−α1 /4C2
2).
The next lemma can be similarly derived as Lemma 2, so its proof is omitted.
18
Lemma 3 Under Assumptions 6-8, taking the bandwidth h =(
logn1
n1
)1/5
, for any δ > 0, there
exists N∗2 such that if n1 ≥ N∗2 ,
P+− (‖E‖max ≥ m) ≤ δ ,
where E is the estimation error matrix as defined in Lemma 1 and m = C3
√2 log(3p/δ)
n1−α1
.
Corollary 1 Under Assumptions 6-8, take the bandwidth h =(
logn1
n1
)1/5
. On the event
J3 ={‖E‖max ≤ C3
√2 log(3p/δ)
n1−α1
}(regarding labeled samples) with P+−(J3) > 1 − δ, there
exists N∗2 ∈ N and C4 > 0 such that if n1 ≥ N∗2 , |Fkl| = |Z1k − Z01k| · |Z1l − Z0
1l| ≤ C4bn1
uniformly for k, l = 1, · · · , p, where bn1 = 2 log(3p/δ)/n1−α1 .
Denote by
J4 =
{32‖E‖max max
S⊂{1,...,p}
|S|φ(S)2
≤ 1
}.
On the event J4, the inequality (4.5) holds, and the compatibility condition is satisfied for Z
(by Lemma 1). Moreover, it can be derived from Lemma 3 by taking a specific δ,
P+−(J4) ≥ 1− 3p exp{−n1−α1 /(2048C2
3A2p)} ,
where Ap = maxS⊂{1,...,p} |S|/φ(S)2. Combining Lemma 2 and the uniform margin condition,
we see that for given estimators f and g, the margin condition holds for the estimated trans-
formed matrix Z involved in the FANS estimator β1. Following similar lines as in van de
Geer (2008) delivers the following theorem, so a formal proof is omitted.
Theorem 1 (Oracle Inequality) In addition to Assumptions 1-8, assume ‖mβ∗1−mβ0
‖∞ ≤η/2 and E(mβ∗
1, f , g)/λ0 ≤ η/4. Then on the event J1 ∩ J2 ∩ J3 ∩ J4, we have
E(mβ1, f , g) + λ‖β1 − β∗1‖1 ≤ 6E(mβ∗
1, f , g) +
16λ2sβ∗1(eη/ε0 + 1)2
cφ21(Sβ∗
1)
.
In addition, when n1 ≥ max(N∗1 , N∗2 ) and under the normalization condition that ‖Z1j‖∞ ≤ 1
for all j, it holds that
P(J1 ∩ J2 ∩ J3 ∩ J4) ≥ 1− exp(−t)− 6p exp{−η2n1−α1 /(4C2
2)} − δ − 3p exp{−n1−α1 /(2048C2
3A2p)} ,
for
λ0 := 4λ∗ +tK
3n+
√2t
n(1 + 8λ∗) ,
where P is the probability with regards to all the samples and
λ∗ =
√2 log(2p)
n+K log(2p)
3n.
19
Theorem 1 shows that with high probability, the excess risk of the FANS estimator can be
controlled in terms of the excess risk of q∗1 when using the estimated density functions f and
g plus an explicit order term. Next, we will study the excess risk of q∗1.
Assumption 9 Let Z01(β1) be the subvector of Z0
1 corresponding to the nonzero components
of β1, and bn1 = log(3p/δ1)/n1−α1 . Assume sβ1
≤ an1 for some deterministic sequence {an1},and an1 · bn1 = o(1). In addition, 0 < C5 ≤ λmin(P{Z0
1(β1)Z01(β1)T}), for some absolute
constant C5.
Assumption 9 allows the number of nonzero elements of β1 to diverge at a slow rate with
n1. Also, it demands a lower bound of the restricted eigenvalue of the sub-matrix of Z0
corresponding to the nonzero components of β1.
Lemma 4 Let Q(β) = Pρ(mβ, f , g) + λ‖β‖0, and β1 = min{|β1,j| : j ∈ Sβ1}. Under
Assumptions 3, 6, 7, 8 and 9, on the event J3, there exists a N∗3 such that, if n1 ≥ N∗3 and
the penalty parameter λ < 0.5C5ε0(1 − ε0)β21, the L0 penalized solution coincides with the
unpenalized version; that is β∗1 = β1.
Theorem 2 (Oracle Inequality) In addition to Assumptions 1-9, suppose 4C1C4s2β0bn1 ≤
λ0η, the penalty parameter λ ∈ (8λ0,min(Lλ0, 0.5C5ε0(1 − ε0) ·minj:β1,j 6=0(|β1,j|))), where C5
is defined in Assumption 9, and ‖mβ∗1−mβ0
‖∞ ≤ η/2. Taking the bandwidth h =(
logn1
n1
)1/5
,
on the event J1 ∩ J2 ∩ J3 ∩ J4 as in Theorem 1, we have
E(mβ∗1, f , g) ≤ C1C4s
2β0bn1 .
Then in view of Theorem 1, we have
E(mβ1, f , g) ≤
16λ2s∗β1(eη/ε0 + 1)2
cφ21(Sβ∗
1)
+ 6C1C4s2β0bn1 .
From Theorem 2, it is clear that the excess risk of the FANS estimator is naturally de-
composed into two parts. One part is due to the nonparametric density estimations while the
other part is due to the regularized logistic regression on the estimated transformed covariates.
When both the penalty parameter λ and the bandwidth h of the nonparametric density esti-
mates f and g are chosen appropriately, the FANS estimator will have a diminishing excess
risk with high probability. Note that one can make explicit λ to obtain a bound on the excess
risk in terms of the sample sizes n and n1, and the dimensionality p.
20
5 Discussion
We propose a new two-step nonlinear rule FANS (and its variant FANS2) to tackle binary
classification problems in high-dimensional settings. FANS first augments the original feature
space by leveraging flexibility of nonparametric estimators, and then achieves feature selection
through regularization (penalization). It combines linearly the best univariate transforms that
essentially augment the original features for classification. Since nonparametric techniques are
only performed on each dimension, we enjoy a flexible decision boundary without suffering
from the curse of dimensionality. An array of simulation and real data examples, supported
by an efficient parallelized algorithm, demonstrate the competitive performance of the new
procedures.
A few extensions are worth further investigation. Beyond a specific procedure, FANS
establishes a general two-step classification framework. For the first step, one can use other
types of marginal density estimators, e.g., local polynomial density estimates. For the second
step, one might rely on other classification algorithms, e.g., the support vector machine, k-
nearest neighbors, etc. Searching for the best two-step combination is an important but
difficult task, and we believe that the answer mainly depends on the data structures.
We can further augment the features by adding pairwise bivariate density ratios. These
bivariate densities can be estimated by the bivariate kernel density estimators. Alternatively,
we can restrict our attention to bivariate ratios of features selected by FANS. The latter has
significantly fewer features.
Dimensions of data sets (e.g., SNPs) in many contemporary applications could be in mil-
lions. In such ultra-high dimensional scenarios, directly applying the FANS (FANS2) approach
could cause problems due to high computational complexity and instability of the estima-
tion. It will be beneficial to have a prior step to reduce the dimensionality in the original
data. Notable works towards this effort on the theoretical front include Fan and Lv (2008),
which introduced the sure independence screening (SIS) property to screen out the marginally
unimportant variables. Subsequently, Fan et al. (2011) proposed nonparametric independence
screening (NIS), which is an extension of SIS to the additive models.
6 Appendix
The appendix contains technical proofs and Lemma 5.
21
Proof of Lemma 2 For any r,m > 0,
P+−(
max1≤j≤p
‖fj − f0,j‖∞ ≥ m
)≤ e−rmE+− exp
(max1≤j≤p
r‖fj − f0,j‖∞)
= e−rmE+−(
max1≤j≤p
exp r‖fj − f0,j‖∞)
≤ e−rmp∑j=1
E+−(
exp r‖fj − f0,j‖∞).
Since we assumed that all fj and f0,j are uniformly bounded by ε−1l , ‖fj−f0,j‖∞ is bounded
by ε−1l for all j ∈ {1, · · · , p}. This coupled with Lemma 1 in Tong (2013), provides a high
probability bound for ‖fj − f0,j‖∞, gives rise to the following inequality,
E+− exp(r‖fj − f0,j‖∞
)≤ exp
r√ log(n1/δ2)
n1h
+ exp(rε−1l ) · δ2 ,
where δ2 playes the role of ε in Lemma 1 of Tong (2013)(taking constant C = 1 for simplicity).
Finding the optimal order for r does not seem to be feasible. So we plug in r = n1−α1 m
and δ2 = exp(−rε−1l ), then
P+−(
max1≤j≤p
‖fj − f0,j‖∞ ≥ m
)
≤ p exp(−n1−α1 m2)
1 + exp
n1−α1 m
√log n1
n1h+n1−α
1 mε−1l
n1h
≤ p exp(−n1−α
1 m2)
1 + exp
√2n1−α1 m
√ log n1
n1h+
√mε−1
l
nα1h
≤ p exp(−n1−α
1 m2)
{1 + exp
[√
2n1−α1 m
(log n1
n1
) 25
+√
2m32 ε− 1
2l n
1110− 3
2α
1 (log n1)110
]},
where in the last inequality we have used the bandwidth h =(
logn1
n1
)1/5
.
The results are derived by taking m =√
2 log(3p/δ1)
n1−α1
( so δ1 = 3p exp(−n1−α1 m2)), and
by taking Assumption 8. Note that we need to introduce α > 0 because the consistency
conditions do not hold for α = 0. In fact, we need at least α > 7/15. Under this assumption,
there exists a positive integer N∗1 such that if n1 ≥ N∗1 ,
1 + exp[2
54 ε− 1
2l n
720− 3
4α
1 (log(3p/δ1))34 (log n1)
110 + 2n
110−α
1 (log(3p/δ1))12 (log n1)
25
]≤ 3 .
22
Therefore, for n1 ≥ N∗1 ,
P+−(
max1≤j≤p
‖fj − f0,j‖∞ ≥ m
)≤ δ1, for m =
√2 log(3p/δ1)
n1−α1
.
Lemma 5 For any vector θ0 = (θ0,1, . . . , θ0,p)T , let Sθ0 = {j : θ0,j 6= 0}, and let the minimum
signal level be θ0 = min{|θ0,j| : j ∈ Sθ0}. Let g(θj) = cj(θj − θ0,j)2 + λ‖θj‖0, where cj > 0. If
λ ≤ cjθ20, g(θj) achieves the unique minimum at θj = θ0,j.
Proof of Lemma 5 For θ0,j = 0, the result is obvious. For θ0,j 6= 0, we have j ∈ Sθ0 and
g(θj) ≥ λ‖θj‖0I(θj 6= 0) + cj(θj − θ0,j)2I(θj = 0)
= λ‖θj‖0I(θj 6= 0) + cjθ20,jI(θj = 0).
If λ‖θ0,j‖0 ≤ cjθ20,
g(θj) ≥ λ‖θj‖0I(θj 6= 0) + λ‖θ0,j‖0I(θj = 0) = λ‖θ0,j‖0.
Since g(θ0,j) = λ‖θ0,j‖0, the lemma follows.
Proof of Lemma 4 Denote Q0(β) = Pρ(mβ, f , g). Then we have β1 = arg minβ∈Rp Q0(β).
Since ∇Q0(β1) = 0 and
∇2Q0(β) = P{Z1ZT
1 exp(ZT
1 β)(1 + exp(ZT
1 β))−2} ≥ ε0(1− ε0)P{Z1ZT
1 } � 0.
By Taylor’s expansion of Q0(β) at β1,
Q(β) = Q0(β1) + 0.5(β − β1)T∇2Q0(β)(β − β1) + λ‖β‖0 , (6.10)
where β lies between β and β1. Let M = P{Z1(β1)Z1(β1)T}, where Z1(β1) is the subvector
of Z1 corresponding to the nonzero components of β1, and M = P{Z01(β1)Z0
1(β1)T}, where
Z01(β1) is the subvector of Z0
1 corresponding to the nonzero components of β1. Let F =
M−M (a symmetric matrix). From the uniform deviance result of Lemma 3, with probability
1 − δ regarding the labeled samples, there exists a constant C4 > 0 such that |Fkl| ≤ C4bn1
uniformly for k, l = 1, · · · , sβ1, where bn1 = 2 log(3p/δ)/n1−α
1 .
Hence, ‖F‖2 ≤ ‖F‖F ≤ C4sβ1bn1 ≤ C4an1bn1 . For any eigenvalue λ(M), by the Bauer-Fike
inequality (Bhatia, 1997), we have min1≤k≤sβ1|λ(M) − λk(M )| ≤ ‖F‖2 ≤ C4an1bn1 , where
λk(A) denotes the k-th largest eigenvalue of A. In addition, in view of Assumption 9, there
exists k ∈ Sβ1such that
λmin(M) ≥ λk(M )− C4an1bn1 ≥ λmin(M )− C4an1bn1 ≥ C5 − C4an1bn1 .
23
Since an1bn1 = o(1), there exists N∗3 (δ) such that when n1 > N∗3 (δ), we have λmin(M ) > 0.
Let β(1)1 be the subvector of β1 consisting of the nonzero components. Then by (6.10) and
Lemma 5 for each j ∈ Sβ1with λ < 0.5C5ε0(1− ε0)β1
2, we have
Q(β) ≥ Q0(β1) + 0.5(C5 − C4an1bn1)ε0(1− ε0)‖β(1) − β(1)1 ‖2 + λ‖β‖0
≥ Q0(β1) +∑j∈Sβ1
{0.5(C5 − C4an1bn1)ε0(1− ε0)(βj − β1,j)
2 + λ‖βj‖0
}, (6.11)
where βj and β1,j are the j-th components of β and β1, respectively. For n1 ≥ N∗3 (δ),
Q(β) ≥ Q0(β1) + λ∑j∈Sβ1
‖β1,j‖0 = Q0(β1) + λ‖β1‖0 .
By (6.10), we have
Q(β1) = Q0(β1) + λ‖β1‖0 .
Therefore, β1 is a local minimizer of Q(β). It then follows from the convexity of Q(β) that
β1 is the global minimizer β∗1 of Q(β).
Proof of Theorem 2 For simplicity, denote by ρ(m(Z1), Y1) the loss function ρq(X1, Y1) =
−Y1m(Z1) + log(1 + exp(m(Z1)). Note that
∂ρ(m(Z1), Y1)
∂m(Z1)= −Y1 +
exp(m(Z1))
1 + exp(m(Z1))= −Y1 + πm,f0,g0(X1),
and
∂2ρ(m(Z1), Y1)
[∂m(Z1)]2=
exp(m(Z1))
[1 + exp(m(Z1))]2.
By the second order Taylor expansion, we obtain that
ρ(mβ(Z1), Y1) = ρ(mβ0(Z0
1), Y1) + [∂ρ(mβ0(Z0
1), Y1)/∂mβ(Z1)](mβ(Z1)−mβ0(Z0
1))
+1
2
∂2ρ(m∗, Y1)
[∂mβ(Z1)]2(mβ(Z1)−mβ0
(Z1))2, (6.12)
where m∗ lies between mβ(Z1) and mβ0(Z0
1). Since
P
[∂ρ(mβ0
(Z01), Y1)
∂mβ(Z1)
]= 0 (6.13)
and 0 < ∂2ρ(m∗, Y1)/[∂mβ(Z1)]2 < 1, taking the expectation we obtain that
|Pρ(mβ(Z1), Y1)− Pρ(mβ0(Z0
1), Y1)| < 0.5P [(mβ(Z1)−mβ0(Z0
1))2]
= 0.5P [(ZT
1 β − (Z01)Tβ0)2].
24
Hence, from Corollary 1, on the event J3,
|Pρ(mβ0(Z1), Y1)− Pρ(mβ0
(Z01), Y1)| ≤ 0.5βT0 P [(Z1 −Z0
1)(Z1 −Z01)T ]β0
≤ C1C4s2β0bn1 ,
where sβ = |Sβ| is the cardinality of Sβ = {j : βj 6= 0}. Naturally, Pρ(mβ0(Z1), Y1) ≤
Pρ(mβ0(Z0
1), Y1) + C1C4s2β0bn1 .
In addition, by definition of β1, Pρ(mβ1(Z1), Y1) = minβ Pρ(mβ(Z1), Y1). As a result,
Pρ(mβ1(Z1), Y1) ≤ Pρ(mβ0
(Z1), Y1). Thus, we have
Pρ(mβ1(Z1), Y1) ≤ Pρ(mβ0
(Z01), Y1) + C1C4s
2β0bn1 . (6.14)
In addition, by (6.12) and (6.13), for any β we have Pρ(mβ(Z1), Y1) ≥ Pρ(mβ0(Z0
1), Y1).
Then, setting β = β1 on the left side leads to
Pρ(mβ1(Z1), Y1) ≥ Pρ(mβ0
(Z01), Y1). (6.15)
Combining (6.14) and (6.15) leads to
|Pρ(mβ1(Z1), Y1)− Pρ(mβ0
(Z01), Y1)| ≤ C1C4s
2β0bn1 . (6.16)
As a result, we have
E(mβ1, f , g) ≤ C1C4s
2β0bn1 . (6.17)
(6.17) combined with Lemma 4 (β∗1 = β1) leads to
E(mβ∗1, f , g) ≤ C1C4s
2β0bn1 . (6.18)
Recall the oracle estimator
β∗1 = arg minβ∈B
{E(mβ, f , g) +
8λsβcφ2
1(Sβ)
}.
Then by Theorem 1,
E(q1) = E(mβ1, f , g) ≤ 6E(mβ∗
1, f , g) +
16λ2sβ∗1(eη/ε0 + 1)2
cφ2(Sβ∗1)
. (6.19)
Therefore, by (6.18) and (6.19),
E(q1) ≤16λ2sβ∗
1(eη/ε0 + 1)2
cφ2(Sβ∗1)
+ C1C4s2β0bn1 .
25
References
Ackermann, M. and Strimmer, K. (2009). A general modular framework for gene set
enrichment analysis. BMC Bioinformatics, 10 1471–2105.
Antoniadis, A., Lambert-Lacroix, S. and Leblanc, F. (2003). Effective dimension
reduction methods for tumor classification using gene expression data. Bioinformatics, 19
563–570.
Bair, E., Hastie, T., Paul, D. and Tibshirani, R. (2006). Prediction by supervised
principal components. J. Amer. Statist. Assoc., 101 119–137.
Bickel, P. J. and Levina, E. (2004). Some theory for Fisher’s linear discriminant function,
’naive Bayes’, and some alternatives when there are many more variables than observations.
Bernoulli, 10 989–1010.
Boulesteix, A.-L. (2004). PLS dimension reduction for classification with microarray data.
Stat. Appl. Genet. Mol. Biol., 3 Art. 33, 32 pp. (electronic).
Buhlmann, P. and van de Geer, S. (2011). Statistics for High-Dimensional Data.
Springer.
Cai, T. and Liu, W. (2011). A direct estimation approach to sparse linear discriminant
analysis. J. Amer. Statist. Assoc., 106 1566–1577.
Clemmensen, L., Hastie, T., Wiiten, D. and Ersboll, B. (2011). Sparse discriminant
analysis. Technometrics, 53 406–413.
Dudoit, S., Fridlyand, J. and Speed, T. P. (2002). Comparison of discrimination meth-
ods for the classification of tumors using gene expression data. J. Amer. Statist. Assoc., 97
77–87.
Fan, J. and Fan, Y. (2008). High-dimensional classification using features annealed inde-
pendence rules. Ann. Statist., 36 2605–2637.
Fan, J., Feng, Y. and Song, R. (2011). Nonparametric independence screening in sparse
ultra-high dimensional additive models. J. Amer. Statist. Assoc., 106 544–557.
Fan, J., Feng, Y. and Tong, X. (2012). A road to classification in high dimensional space:
the regularized optimal affine discriminant. Journal of the Royal Statistical Society. Series
B (Statistical Methodology), 74 745–771.
26
Fan, J. and Li, R. (2001). Variable selection via nonconcave penalized likelihood and its
oracle properties. J. Amer. Statist. Assoc., 96 1348–1360.
Fan, J. and Lv, J. (2008). Sure independence screening for ultrahigh dimensional feature
space (with discussion). J. Roy. Statist. Soc., Ser. B: Statistical Methodology, 70 849–911.
Friedman, J., Hastie, T., Hofling, H. and Tibshirani, R. (2007). Pathwise coordinate
optimization. Annals of Applied Statistics, 1 302–332.
Friedman, J., Hastie, T. and Tibshirani, R. (2010). Regularization paths for generalized
linear models via coordinate descent. Journal of Statistical Software, 33 1–22.
Gordon, G. J., Jensen, R. V., Hsiao, L.-L., Gullans, S. R., Blumenstock, J. E.,
Ramaswamy, S., Richards, W. G., Sugarbaker, D. J. and Bueno, R. (2002).
Translation of microarray data into clinically relevant cancer diagnostic tests using gene
expression ratios in lung cancer and mesothelioma. Cancer Research, 62 4963–4967.
Guo, Y., Hastie, T. and Tibshirani, R. (2007). Regularized linear discriminant analysis
and its application in microarrays. Biostatistics, 8 86–100.
Hastie, T. and Tibshirani, R. (1990). Generalized Addtive Models. Chapman & Hall/CRC.
Hastie, T., Tibshirani, R. and Friedman, J. H. (2009). The Elements of Statistical
Learning: Data Mining, Inference, and Prediction (2nd edition). Springer-Verlag Inc.
Huang, P. W., X. (2003). Linear regression and two-class classification with gene expression
data. Bioinformatics, 19 2072–2978.
Li, K.-C. (1991). Sliced inverse regression for dimension reduction. J. Amer. Statist. Assoc.,
86 316–342. With discussion and a rejoinder by the author.
Mai, Q., Zou, H. and Yuan, M. (2012). A direct approach to sparse discriminant analysis
in ultra-high dimensions. Biometrika, 99 29–42.
Meier, L., Geer, V. and Buhlmann, P. (2009). High-dimensional additive modeling.
Ann. Statist, 37 3779–3821.
Nguyen, D. V. and Rocke, D. M. (2002). Tumor classification by partial least squares
using microarray gene expression data . Bioinformatics, 18 39–50.
Polonik, W. (1995). Measuring mass concentrations and estimating density contour clusters-
an excess mass approach. Annals of Statistics, 23 855–881.
27
Raykar, V., Duraiswami, R. and Zhao, L. (2010). Fast computation of kernel estimators.
Journal of Computational and Graphical Statistics, 19 205–220.
Shao, J., Wang, Y., Deng, X. and Wang, S. (2011). Sparse linear discriminant analysis
by thresholding for high dimensional data. Ann. Statist., 39 1241–1265.
Tibshirani, R. (1996). Regression shrinkage and selection via the lasso. J. Roy. Statist.
Soc., Ser. B, 58 267–288.
Tong, X. (2013). A plug-in approach to anomaly detection. Journal of Machine Learning
Research, 14 3011–3040.
Tsybakov, A. (2004). Optimal aggregation of classifiers in statistical learning. Ann. Statist.,
32 135–166.
van de Geer, S. (2008). High-dimensional generalized linear models and the lasso. Ann.
Statist., 36 614–645.
Witten, D. and Tibshirani, R. (2012). Penalized classification using fisher’s linear dis-
criminant. Journal of the Royal Statistical Society Series B, 73 753–772.
Wu, M. C., Zhang, L., Wang, Z., Christiani, D. C. and Lin, X. (2009). Sparse linear
discriminant analysis for simultaneous testing for the significance of a gene set/pathway
and gene selection. Bioinformatics, 25 1145–1151.
Zhang, C.-H. (2010). Nearly unbiased variable selection under minimax concave penalty.
Ann. Statist., 38 894–942.
Zhang, C.-H. and Huang, J. (2008). The sparsity and bias of the LASSO selection in
high-dimensional linear regression. Ann. Statist., 36 1567–1594.
Zhao, P. and Yu, B. (2006). On model selection consistency of lasso. Journal of Machine
Learning Research, 7 2541–2563.
Zou, H. (2006). The adaptive lasso and its oracle properties. J. Amer. Statist. Assoc., 101
1418–1429.
Zou, H., Hastie, T. and Tibshirani, R. (2006). Sparse principal component analysis. J.
Comput. Graph. Statist., 15 265–286.
28